How to Master AI-Assisted Coding: A Step-by-Step Guide to Agentic Engineering

Introduction

Artificial intelligence is transforming how we write software, but the real challenge isn't generating code—it's verifying that the code works. Chris Parsons, a leading voice in AI engineering, has updated his practical guide on using AI for coding, emphasizing a shift from blind trust to systematic validation. Based on his insights and those of experts like Simon Willison and Birgitta Böckeler, this step-by-step guide will help you adopt agentic engineering—a disciplined approach where AI generates solutions and you build harnesses to verify them. Whether you're a senior developer or a team lead, these steps will help you move from 'vibe coding' (ignoring the output) to a rigorous, replicable workflow.

What You Need

• AI coding tool: Claude Code or Codex CLI (recommended by Parsons for their inner harness features)
• Version control system: Git or equivalent for tracking changes
• Automated testing framework: e.g., Jest, pytest, or your language's standard testing tool
• Type checker or linter: TypeScript, mypy, ESLint, etc.
• Continuous integration (CI) pipeline: For running automated gates
• Documentation tool: Markdown, Confluence, or a wiki
• Team communication channel: Slack, Teams, or similar

Step-by-Step Guide

Step 1: Choose Your AI Tool and Set Up the Inner Harness

Start by selecting one of the recommended tools—Claude Code or Codex CLI. Both provide an 'inner harness' that allows you to define guardrails and automated checks within the AI interaction. Configure the tool to run tests, type checks, and lints automatically after each code generation. This aligns with Parsons' advice: "Build better review surfaces, not better prompts." Set up a project environment that mirrors production as closely as possible so the AI can verify against realistic conditions before asking for human input.

Step 2: Keep Changes Small and Incremental

Break each task into the smallest possible unit of work. For each piece, instruct the AI to generate only a few lines or one function. This makes verification manageable and reduces the risk of cascading errors. As Parsons notes, a team that can generate five approaches and verify them in an afternoon outperforms one that waits a week for feedback. Small changes also make it easier to roll back if something goes wrong.

Step 3: Build Automated Guardrails

Create a set of automated gates that must pass before any code can be merged. These include unit tests, integration tests, type checking, and static analysis. Use the tool's inner harness to run these checks immediately after generation. For example, if the AI suggests a new API endpoint, automatically run the existing test suite against it. The goal is to make verification non-negotiable and to catch issues before they reach human review. As Parsons says, "The check still happens; it just doesn't always happen in your head."

Step 4: Document Ruthlessly

Add comments, documentation strings, and inline explanations for every AI-generated change. This not only helps future developers but also trains the AI to produce better code over time. When the AI sees that you expect clear documentation, it will begin to include it automatically. Use a consistent style (e.g., JSDoc, Sphinx) and enforce it through the linting rules in your harness. Parsons insists that documentation is a fundamental part of the agentic engineering workflow—without it, the 'speed' of AI becomes a liability.

Step 5: Verify Every Change Through Automated Gates

Implement a policy: no code is merged until it passes all automated checks AND is reviewed by a human (or by a senior engineer for critical decisions). However, shift the balance toward automation. The phrase 'verified' now means "checked by tests, by type checkers, by automated gates, or by you where your judgement matters." Use CI to run the full harness on every commit. If a test fails, the AI must revise its output before resubmitting. Teach the AI to read error messages and self-correct—this is the core of training.

Step 6: Train the AI by Providing Clear Feedback

When the AI's output passes all gates but still seems suboptimal (e.g., poor performance, unnecessary complexity), provide explicit feedback in the prompt or through the harness. For example: 'The generated code doesn't handle edge case X. Rewrite with a guard clause.' Over time, the AI learns your preferences. This is where senior engineers add the most value—shaping the AI's behavior so that future diffs are right the first time. Parsons emphasizes that "the most important thing skilled agentic programmers can do is pass that skill onto other developers." So document your feedback patterns and share them.

Step 7: Evolve from Reviewer to Harness Shaper

Instead of spending your days approving diffs, invest time in improving the harness itself. Add new tests, update static analysis rules, create environment simulations, and expand the automated verification suite. This is where your expertise compounds. As Parsons writes, "The way out is to train the AI so the diffs are right the first time, to make yourself the person on the team who shapes the harness, and to make that work the visible thing you are measured on." That role is far more valuable than reviewing low-level code changes.

Step 8: Invest in Fast Feedback Cycles

Optimize your pipeline so that the AI can get feedback in seconds, not hours. Use parallel test execution, cloud runners, and caching. Implement 'computational sensors'—as Birgitta Böckeler calls them—such as static analysis and quick smoke tests that run directly in the AI's environment. The faster the cycle, the more iterations you can afford, and the better the final result. A team that verifies five approaches in an afternoon will outpace one that verifies one in a week.

Tips for Success

• Start with a small pilot project. Apply these steps to a non-critical codebase first to refine your harness.
• Use computational sensors early. Static analysis and quick tests are cheap to run and catch many common errors.
• Measure what matters. Track how often the AI passes all gates on the first attempt—that's your proxy for training effectiveness.
• Embrace the shift in your role. If you're a senior engineer, your primary output is no longer code—it's the harness and the trained AI.
• Learn from others. Birgitta Böckeler's video with Chris Ford on harness engineering is a great resource for deeper insights.
• Revisit your setup monthly. AI tools evolve quickly; update your harness to take advantage of new capabilities.